Memory enables LLM‑based multi‑agent systems to reuse past experience, yet stored advice can be erroneous or applied in unsuitable situations. Relying solely on a team’s overall success or failure does not reliably indicate whether a memory should be corrected or discarded. To address this, we introduce DiscernMem, a training‑free framework that simultaneously improves memory quality and guides its proper use.\ \ Defect‑Aware Memory Refinement (DMR) compares the execution of a task with and without a given memory, diagnoses rule defects, and tests repairs on both the current and earlier tasks to prevent loss of previously useful behavior.\ \ Boundary‑Learning Meta‑Memory (BLM) logs each memory’s past usages and outcomes, learning when the owning role should apply or avoid the memory based on task context and incoming messages, thereby adapting the boundaries of memory usage.\ \ Experiments with Qwen3-4B-Instruct-2507 across four benchmarks show that DiscernMem achieves the highest average scores on AutoGen, CAMEL, and MacNet, surpassing the strongest baselines by 0.51‑4.32 points. It leads in both code generation and symbolic planning, attaining a PDDL score of 37.71 on CAMEL versus 28.48 for the next best method.\ \ Review